AutoGPT's AutoPilot now supports a local transport (`CHAT_USE_LOCAL`) that routes chat to a self-hosted OpenAI-compatible endpoint (e.g., Ollama, vLLM) instead of requiring an OpenRouter or Anthropic API key, with runtime context-window detection and compaction fixes to prevent prompt corruption.
Adds a local transport profile to AutoGPT's AutoPilot chat that routes requests to a self-hosted LLM backend without API keys, auto-detects the backend's context window at runtime, and fixes a compaction bug that previously caused prompt corruption on small-window models.
Developers and operators can now run AutoGPT's AutoPilot entirely on local hardware without any cloud API keys, making self-hosted experimentation and air-gapped evaluation possible. The runtime context-window probe and compaction fix prevent the silent prompt corruption that occurred when the old 120k default exceeded a small local model's window, so conversations no longer lose the system prompt mid-session. Watch for stability on macOS reboots (Ollama service persistence is out-of-scope) and for edge cases where the probe returns a stale cached window after server restarts.